Potential Underestimation of Satellite Fire Radiative Power Retrievals over Gas Flares and Wildland Fires

Abstract

Fire Radiative Power (FRP) is related to fire combustion rates and is used to quantify the atmospheric emissions of greenhouse gases and aerosols. FRP over gas flares and wildfires can be retrieved remotely using satellites that observe in shortwave infrared (SWIR) to middle infrared (MIR) wavelengths. Heritage techniques to retrieve FRP developed for wildland fires using the MIR 4 μm radiances have been adapted for the hotter burning gas flares using the SWIR 2 μm observations. Effects of atmosphere, including smoke and aerosols, are assumed to be minimal in these algorithms because of the use of longer than visual wavelengths. Here we use Moderate Resolution Imaging Spectroradiometer (MODIS), Visible Infrared Imaging Radiometer Suite (VIIRS) and Landsat 8 observations acquired before and during emergency oil and gas flaring in eastern Saudi Arabia to show that dark, sooty smoke affects both 4 μm and 2 μm observations. While the 2 μm observations used to retrieve gas FRP may be reliable during clear atmospheric conditions, performance is severely impacted by dark smoke. Global remote sensing-based inventories of wildfire and gas flaring need to consider the possibility that soot and dark smoke can potentially lead to an underestimation of FRP over fires.

1. Introduction

Atmospheric emissions of greenhouse gases and aerosols from fires have substantial climate-forcing effects [1], and there is a need for global monitoring. Operational remote sensing-based methodologies have been developed to retrieve fire radiative power (FRP), and have been used in applications to estimate combustion/emission rates over at least the last two decades. More recently, FRP has been shown to be useful for understanding the effect of fire on plant physiology and mortality [2,3,4]. FRP can be retrieved remotely using single band radiance observations using approximations to Planck’s law over a finite range of temperatures [5,6,7] and assuming clear observing conditions. Using this technique, Fisher et al. [5] derived and demonstrated a method for the remote retrieval of FRP from hot (>1600 K) gas flares using shortwave infrared (SWIR) 1.6 μm and 2 μm wavelength nighttime radiance observations, based on established [6,7] methods for relatively cooler (400–1000 K) wildfire FRP measurements, which use middle infrared (MIR) 4 μm radiance observations. The effect of atmospheric smoke and aerosols, although known to attenuate [6,8,9] the signal arriving at the sensor, are largely unconstrained, and have been assumed to be minimal [5,8,10,11,12,13,14] at longer visual wavelengths.

To the best of our knowledge, the attenuation of FRP by soot and dark smoke has not been systematically studied. Such studies are important to reduce the uncertainties in global greenhouse gas emission estimates from both biomass and gas flares. Further, FRP metrics can be used to estimate changes in net primary productivity (NPP) as an estimate of vegetation damage (i.e., burn severity) of wildfire and prescribed fire [3]. Monitoring burn severity using FRP instead of common spectral metrics (e.g., normalized burn ratio and derivatives), could result in better predictions of post-fire tree growth and damage for land managers [4]. Given reported increases in the number of severe wildfire events, particularly in the United States [15,16], the application of FRP will likely continue to increase. If smoke and soot limitations to FRP are not addressed, estimates of ecophysical change could be biased.

The difficulty of obtaining independent reference data at the time of the satellite overpass makes the validation of active fire products challenging, because both wildland fires and atmospheric conditions change continuously in space and time [17,18]. Due to their well-constrained locations, active gas flares have been used for the characterization and accuracy assessments [9,19] of active fire products. Gas flares are frequently detected by moderate spatial resolution satellite platforms, such as the 30 m Landsat [14] and 375 m Visible Infrared Imaging Radiometer Suite (VIIRS), as detection rates depend upon the relative size of the fire to the detection footprint. Coarse sensors, such as the 1 km Moderate Resolution Imaging Spectroradiometer (MODIS) satellite, detect relatively larger gas flares.

Recently, the news media (e.g., [20,21,22]) reported emergency oil and gas flaring over four oil fields near Abqaiq, Saudi Arabia. The extensive dark sooty smoke seen in satellite imagery since 14 September 2019 triggered environmental and health concerns. Media reports suggested that the dark smoke was the result of emergency flaring [13], an established method to safely dispose of the excess oil and gas that a plant cannot process. This event provided a unique opportunity to compare satellite observations before and during the emergency flaring to understand the effect of smoke and soot on remote FRP retrievals.

In this letter, we compare visual (VIS) to MIR wavelength observations retrieved from MODIS, VIIRS, and Landsat 8 satellites, which have different overpass times and nominal spatial resolutions ranging from 30 m to 1 km, before and during emergency flaring. It is noted that gas flare FRP retrievals at 2 μm are generally retrieved using nighttime observations [5,23]; however, we use the daytime observations to qualitatively illustrate the effect of dark smoke readily identified in daytime imagery. Higher 4 μm FRP and 2 μm observations were expected during the emergency flaring event compared to pre-flaring observations. We found that although the 4 μm FRP measurements are higher than pre-flaring levels, the 2 μm band observations are mostly lower. A high degree of variability in the 4 μm FRP among the satellites also was seen. These results indicate that SWIR ~2 μm and shorter wavelength observations are impacted more by the presence of dark smoke and may underestimate gas FRP under sooty conditions. Dark, sooty smoke typically results from incomplete combustion, and can also appear over fires with limited oxygen supplies, such as in smoldering fires. Thus, understanding the effects of soot may help better constrain atmospheric emissions from such gas flares and wildland fires.

2. Data and Methods

2.1. VIS- SWIR 2 μm Data

Landsat 8 is a polar-orbiting satellite that can image the globe in 11 spectral bands, ranging from reflective to emissive thermal wavelengths, at a spatial resolution of 30 and 100 m, respectively, and includes a 15-m panchromatic band. The Landsat 8 satellite has a revisit frequency of 16 days, with a required nominal equatorial crossing mean local time of 10:00 AM ± 15 min [24]. The Landsat 8 onboard Operational Land Imager (OLI) instrument measures VIS-SWIR radiance/reflectance data in seven spectral bands, and is produced in approximately 185 × 180 km tiles defined in a Worldwide Reference System (WRS-2) [24] of paths and rows. The most recent and available calibrated and terrain-corrected top of atmosphere L1T product includes Landsat 0.65 μm band 4, 0.56 μm band 3, 0.48 μm band 2 and 2.20 μm band 7 reflectance.

Landsat 8 data acquisitions over the Saudi Arabia fires (WRS-2 path 164 and row(s) 043 and 042) that were acquired on the 1st and 17th of September, are used in this study for qualitative illustration only. Landsat 8 data were downloaded from the National Aeronautical and Space Administration (NASA) Land Processes Distributed Active Archive Center LPDAAC [25].

MODIS sensors on board the Aqua and Terra polar-orbiting platforms can map the Earth in 36 spectral bands, spanning reflective to thermal wavelengths, up to two times each day and night, with nominal overpass times at local 10:30 AM/PM and 1:30 AM/PM at the equator. MODIS Level-3 data are gridded in horizontal (h) and vertical (v) tiles, and the 500-m reflectance data are available as the MYD09GA (Aqua) and MOD09GA (Terra) standard collection 6 products. These products include calibrated and gridded 500 m 2.13 μm surface reflectance. Daily MODIS SWIR data from the 1st to the 21st of September were downloaded from the NASA Land Processes Distributed Active Archive Center (LPDAAC) [25].

2.2. MIR 4 μm FRP Data

MODIS and VIIRS sensors also have the MIR 4 μm band required to detect fires that are actively burning, and retrieve FRP at the time of overpass. The 375-m VIIRS sensor is on board the Suomi National Polar-Orbiting Partnership (Suomi NPP) spacecraft with equatorial over pass around local 1: 30 AM/PM. We use the VIIRS VNP14IMGTDL and collection 6 MODIS Level 3 Aqua (MYD14) and Terra (MYD14) satellite active fire products in this study. Each 375-m (VIIRS) or 1 km (MODIS at nadir) observation identifies whether an active fire was detected and retrieves the FRP in megawatts MW. Active fires are detected by their high and contrasting 4 μm brightness temperatures compared to the neighboring pixels. A series of locally adaptive contextual tests using 4 μm observations and other wavelengths are also used to reject false alarms. Once a fire pixel is detected, its FRP is derived using a linear relationship between the FRP and the 4 μm radiance [7]. The contribution of non-burning background radiance to the FRP is minimized [6,7] by subtracting the background 4 μm radiance from the fire pixel. Algorithms specific for MODIS [6,7] and VIRS [26] FRP have been developed using the MIR bands [6,7], and are being operationally used. Active fire data from both MODIS and VIIRS between the 1st September and 21st of September 2019 for MODIS tile h22; v06 was downloaded from the NASA Fire Information for Resource Management System (FIRMS) [27].

2.3. Methods

We compare the 2 μm and 4 μm satellite observations before and after the day of emergency flaring over the four oil fields at Abqaiq, Saudi Arabia, reported by [20,21,22]. A unique spatial extent box is conservatively defined for each of the four sites for analysis as the smallest rectangular extent that encompasses all central locations of active fire detections from both MODIS and VIIRS satellites over each site irrespective of the day or time of detection between the 1st and 21st of September 2019. The smallest rectangular box enables a consistent spatial extent for analysis on all days spanning before and during emergency flaring. Daily time series summaries of 2 μm and 4 μm satellite observations over each flare specific extents are presented and discussed. Coherent increases and decreases in observations across SWIR and MIR wavelengths are highlighted and discussed considering the influence of dark, sooty smoke.

3. Results

Thick, black smoke emanating from the emergency oil gas flaring sites not visible on the 1st of September 2019 is evident in the Landsat 8 images in both VIS and SWIR band 7 reflectance (Figure 1) over four locations in the image acquired on 17th September 2019. The SWIR band is not totally immune to the spectrally disruptive effects of thick, black smoke, although it is known to be less sensitive to smoke and aerosols compared to visible wavelengths.

To investigate the effect of soot and smoke further, the active fire detections in a time period that encompassed the day of flaring were examined. Figure 2 illustrates the active fire detection locations by VIIRS and both MODIS Aqua and Terra platforms between the 1st and the 21st of September. The background image in Figure 2 is from the 17th of September. It is clear each site may have multiple gas flares, and that smoke can reduce the observed reflectance, and may even obscure gas flares. To study the effect of smoke on FRP, the time series of FRP retrievals is examined.

Figure 3 summarizes the time series of FRP as observed over the four distinct emergency flaring locations. VIIRS detected active fires over these locations on most days before 14th September 2019, while MODIS detections are predominantly after the flaring. This is expected, as detection probabilities are proportional to the fractional area covered by fire with respect to the pixel footprint.

Hence, the smaller (375-m) pixel dimension of VIIRS has increased the probabilities of detection as compared to MODIS 1 km. FRP retrieved from both MODIS Aqua and Terra and VIIRS are summarized as box and whisker plots in Figure 3. A distinct increase in the FRP is seen after the 14th of September 2019 over all sites by all satellites as compared to pre-flare FRP, which is expected due to continued active flaring at least until the end of the study period, which was the 21st of September, 2019. It is also evident that the three satellite observations differ considerably over these four locations during the time of emergency flaring. Although the gas stacks themselves cover a small fraction of the MODIS and VIIRS pixels, flame size, atmospheric effects and sensor characteristics such as the point spread function can all influence the possibility of detections over multiple [9] pixels for the same subpixel fire. Different sensors, overpass times, view angle and atmospheric effects can influence FRP retrievals.

The increased FRP is highly variable with distinct decreases relative to FRP on the 14th of September on certain dates, specifically on the 16th and on the 21st of September over all sites. The increase in FRP was expected after the emergency flaring, but the decrease on certain days can be explained only if flare rates were all simultaneously reduced on all four specific sites, or if they had similar viewing conditions that resulted in reporting reduced FRP. To investigate this further, the daily 2 μm observations over the sites for the same four locations were analyzed.

Figure 4 summarizes the observed MODIS Aqua and Terra 2.13 μm reflectance over the flaring sites, irrespective of the day or time of detection between the 1st and the 21st of September 2019. Significant increase in 4 μm FRP and 2 μm reflectance over the flare is expected if the effect of smoke was negligible. However, it is evident that 2 μm reflectances are mostly low, with very few isolated observations that were higher on certain days than pre-emergency flaring days (Figure 4). The low 2 μm reflectance values strongly suggest the attenuating effect of dark smoke. On the 17th and 21st of September, the reflectance values are all lower, thereby corresponding with the lows in FRP (Figure 3). This simultaneous decrease in signal can happen if the flaring rate was decreased on those days, or can be an artifact of dark smoke. The Landsat 8 image (Figure 1 and Figure 2) on the 17th of September 2019 clearly indicates heavy, thick smoke and continued flaring, which strongly suggests that the attenuation of FRP and the 2 μm observations result from sooty smoke. The presence of soot and subsequent lowering of the 2 μm reflectance continue to be observed (Figure 4) after the flare date and throughout the study period. The lowering of fire signal in a pixel during the day time is the result of dark smoke obscuring both the fire and adjacent non-burning land in the pixel footprint. However, during nighttime, dark smoke obscuring of the flame may lead to an attenuation of the signal and hence the inferred FRP.

4. Conclusions

We compared the MIR 4 μm FRP observations from MODIS, VIIRS, before and during the emergency flaring alongside the SWIR 2 μm observations using daily MODIS Aqua and Terra and available Landsat 8 observations. We show that the presence of smoke can lower 4 μm FRP retrievals and 2 μm signals, especially over emergency flaring gas sites that produce dark smoke. These results suggest that global remote sensing-based gas flare inventorying programs may potentially underestimate emissions [5,23]. Given that a variety of factors may lead to emergency oil and gas flaring, and that emergency flaring may not always be reported [13], it is important to consider the possibility of flares going under-reported or undetected through current remote sensing methods.

The production of dark smoke is usually an indicator of incomplete combustion in gas flares, as well as in wildland fires with an inadequate oxygen supply (i.e., smoldering fires). Smoldering fires are the most emissive and have the lowest probability of detection [14,17] due to smoke and lower temperatures. Global gas flaring inventory [13,23] methodologies and wildfire monitoring need to recognize consider that soot can lower detection rates and underestimate FRP retrievals. Sooty gas flares may be identified by monitoring the 4 μm FRP, which is affected less by smoke and soot. Smoke obscuration is an issue for remote wildland fire detections using satellites [9,18] as well. Future studies can advance this work by exploring the use of wavelengths longer than MIR to flag smoky observations and constrain emissions from smoldering fires. Laboratory studies may help identify methodologies for the detection and quantification of FRP underestimation resulting from soot, as well. The growing number of sensors with active fire capabilities exploiting different spectral and temporal observations should also help constrain the effect of smoke for both gas flares and wildland fires.